The Humanoid Robot Story Was Never About the Robot

The humanoid robot story was never about the robot. It's about who owns the liability layer underneath it, and almost nobody covering this beat is writing about that part.

Start with what's actually being reported this month. Figure's robots have contributed to production of more than 30,000 BMW X3 vehicles at the Spartanburg plant, running material handling and parts transfer on ten-hour shifts, five days a week. It's the most documented commercial humanoid deployment in the industry right now. The newer Figure 03 has since moved to a different job at the same plant: sorting unsorted components into sequencing trolleys for just-in-time delivery to the line, work a person used to do by hand. AgiBot, the Chinese humanoid maker, claims roughly 15,000 cumulative units shipped: a company figure, not an audited one. LG is testing wheel-based CLOiD robots inside a live washing machine factory in Tennessee, specifically to see how the platform behaves in an operating environment rather than a lab. And Hyundai, which owns Boston Dynamics and its decades of hydraulic and dynamic-balance research behind the Atlas platform, says it won't deploy humanoid robots at its US factories until 2028.

That spread should bother you more than it does. If this were purely a hardware maturity story, you wouldn't see one company running robots on a production line today and another company, with arguably better robotics assets, pushing deployment out two more years. Something other than actuator quality is setting the pace.

The Number That's Missing

Here's the detail that gives it away. BMW's own press materials on the Figure 03 logistics deployment don't disclose payload capacity, cycle rate, robot count, or commercial terms. That's not an oversight from a company that publishes detailed specs on everything else it puts on a factory floor. Companies disclose uptime and unit economics once the deal is proven and the liability terms are settled. The numbers become a selling point. They stay quiet while the insurance and the audit trail are still being negotiated behind the press release.

Liability is the constraint.

The First Precedent

This isn't a new problem. It's an old one wearing a new chassis. The first industrial robot, Unimate, went to work at a GM die-casting plant in Trenton in 1961, lifting hot metal parts too dangerous for a person to handle by hand. Engelberger, the man who commercialized it, built the sales pitch around exactly this: put the robot where a human could lose a limb, and take the human out of that space entirely. The safety solution that followed for the next sixty years was structural, not computational: cage the robot, fence the work envelope, keep the two apart. ISO 10218, the standard that still governs industrial robot safety, only added serious provisions for robots and humans sharing space in its most recent editions, because for decades the entire discipline assumed they wouldn't.

That worked because the fence did the liability's job for it. A caged robotic arm doesn't need to explain why it moved a certain way at a certain moment. Nobody was close enough to be hurt by the answer. Humanoid robots don't get the fence. Figure 03 sorting parts into trolleys next to a human line worker, or an LG CLoiD robot sharing floor space in an active plant, has no cage standing between the machine and the liability question. Every incident, even something as small as a dropped part, needs an answer to why the robot did what it did, one attributable enough to satisfy an insurer, and eventually a plaintiff's attorney. That answer doesn't exist yet as a standardized product. It's being built ad hoc, deployment by deployment, inside deals like BMW's. That's exactly why the terms aren't public.

The Closer Precedent

Autonomous vehicles already ran this experiment. Here's how it actually went. Cruise's robotaxis operated in San Francisco for years on a straightforward premise: the driving problem was mostly solved, scale was the remaining work. Then in October 2023, a Cruise vehicle dragged a pedestrian roughly 20 feet after an initial collision (the vehicle tried to pull over with her still trapped underneath), and California's DMV suspended Cruise's entire driverless permit within days. The underlying self-driving technology hadn't gotten worse. Cruise had withheld the full video, and the regulator no longer trusted the company's account of what its own system did and why. GM shut Cruise's robotaxi operations down entirely within the year.

The lesson wasn't "self-driving doesn't work." Cruise's cars had logged enormous mileage without incident before that night. The lesson was that the entire business lived or died on one capability that had nothing to do with driving: producing a fast, trustworthy account of why the machine did what it did, the moment someone in authority asked. Cruise had the actuators. It didn't have the audit trail, or at least didn't have the institutional habit of surfacing it immediately and completely. That gap is what ended the company's robotaxi business.

The Skeptical Read

The obvious counter: this is pilot theater, not a liability story. BMW ran Figure 02 for eleven months and is now running Figure 03 on a different task. That reads like a company still evaluating, not one that solved a liability problem and is scaling. Hyundai pushing to 2028 could just mean Boston Dynamics' hardware isn't ready, full stop, no liability subplot required. And AgiBot's 15,000-unit figure is a single-sourced company claim with no independent audit behind it: treat it as marketing until proven otherwise.

That's a legitimate read, and the AgiBot number specifically should be flagged as unverified. But it doesn't explain the specific pattern of disclosure. Pilot-stage companies that are purely hardware-constrained tend to publish benchmarks (cycle times, error rates, uptime, unit costs) because that's the story that attracts the next funding round or OEM contract. What we're seeing instead is companies publishing outcomes ("supported 30,000 vehicles") while withholding mechanisms (payload, cycle rate, count, terms). That's the disclosure pattern of a legal and insurance negotiation still in progress, not an engineering one.

Who Writes the Rules This Time

Nobody has written the humanoid-robot equivalent of ISO 10218 yet, because the caged-robot assumption baked into that standard doesn't apply to a machine sorting parts next to a person. OSHA has no humanoid-specific standard on the books. Insurers underwriting these deployments today are doing so case by case, pricing risk off deployment-specific data because there's no actuarial baseline yet: no accumulated loss history the way there is for, say, forklifts or fixed robotic arms. That absence is exactly why BMW's terms with Figure aren't public. There's no industry-standard contract to point to. Every one of these deals is bespoke because the rulebook doesn't exist yet, which means whoever writes a rulebook the rest of the industry actually adopts gets to set the terms everyone else negotiates against.

That's a bigger prize than it sounds. Standards bodies and certification labs don't usually get remembered as the interesting part of a technology transition. But in every prior wave (aviation, elevators, food safety, financial data), the entity that wrote the standard other people had to comply with ended up sitting permanently upstream of the operators who did the visible work. Physical AI doesn't have that entity yet. Someone is going to become it.

Where the Value Actually Accrues

If liability is the real bottleneck, the value doesn't accrue to whoever builds the best actuator. It accrues to whoever builds the layer that makes a physical AI system's decisions auditable, attributable, and insurable at the moment something goes wrong. Call it the physical-world equivalent of compliance-as-code. That's a categorically different business than humanoid robot manufacturing: it looks more like an insurance underwriter or a safety-certification body than a robotics company. Whoever builds it first doesn't sell one robot at a time. They sell the terms every OEM deployment has to run through.

Concretely, that layer has to do two things Cruise's own systems apparently couldn't do fast enough: log every decision a physical system makes in a form that survives the incident, and produce that log to a regulator or insurer, without a company's PR function in the loop, in a consistent, auditable format across every robot vendor and every factory floor. Nobody has productized that yet. Right now it's a clause in an unpublished BMW contract, not a company you can invest in. The first team that turns it into the latter owns the toll booth every humanoid deployment after them has to pass through.

The Asymmetric Implication

The market is pricing this as a hardware race: whoever builds the most capable, most cost-effective humanoid robot wins. That's the wrong unit of analysis. The actual race is for the liability layer sitting underneath every one of these deployments, and right now it's being built privately, contract by contract, inside deals like BMW's rather than as a standalone product anyone can buy. Whoever productizes it (turns bespoke liability negotiation into a standard, auditable layer any OEM can plug a physical AI system into) captures a disproportionate share of the value every humanoid deployment after them creates.

The robots are the visible part. The audit trail is the business.